Engineer swaps LLM explanations for playable simulations

Laurentiu Raducu turns technical topics into playable low-poly simulations built with Claude and OpenCode, according to his blog post that topped Hacker News with 680 points and 433 comments.

The workflow flips the usual pattern for learning from an LLM. Instead of asking for an explanation, you ask for something to break. He's built five so far. ChipTycoon covers chip fabrication, there's a rocket engine sim, Token Town explains LLM internals, plus an F1 power unit sim and one for EUV lithography. Each one ships with real controls tied to the underlying mechanics.

The process runs in four steps. First, establish the fundamentals in plan mode, using Claude or OpenCode to sketch what the simulation needs to model. Second, have the model audit its own explanation for accuracy before any code gets written. Third, build the sim itself, a RollerCoaster Tycoon-style interface with sliders and levers instead of paragraphs. Fourth, deploy it to GitHub Pages, so people can play it instead of read about it.

Raducu calls the results "100% accurate and free of hallucinations." That claim doesn't hold up. LLMs don't verify themselves into certainty because a second pass ran. The workflow still works without it.

The real value is the audit step, not the accuracy claim. Skip it, and the simulation only launders whatever the model got wrong into something that feels true because you can click on it. Watch whether Raducu or anyone building on this publishes a case where the audit step actually caught an error. That's the evidence the workflow needs.